Optical holographic devices and systems present an emerging technology for information storage and processing devices and systems. Optical holographic devices and systems have interesting properties such as high speed in data transfer and processing, the capability of massive parallel processing, and superior resistance to electromagnetic interference. One unique property of holographic memory devices is their high capacity data storage, fast access time and transfer rate. This is particularly useful in applications that demand such high capacity. Other examples of optical holographic systems include holographic correlators for image recognition, holographic associative memories, holographic neuromorphic systems, and holographic interconnection systems.
Certain drawbacks, however, have limited the use of many optical holographic devices and systems. For example, imaging optics is often used in many prior-art systems. This places limit on miniaturization of these systems. The optical alignment of an imaging assembly is often stringent and sensitive to factors such as vibration and temperatures. The limitations in the prior-art optical systems become significant and often present obstacles to the full use of the unique advantages that are inherent in optical holographic systems.
Two conventional systems are examined in the following to further illustrate the advantages of optical holographic systems as well as the limitations of conventional optical implementations.
FIG. 1 is a schematic of a conventional holographic system 100 for three dimensional data storage. Information is stored in the holographic medium 102 in the form of multiple superimposed holograms. Each hologram is recorded by the interference of a signal beam 110 coming from the signal arm 104 with a reference beam 120 from the signal arm 104. A transmissive spatial light modulator (SLM) 112 in the reference arm 106 imprints data on the signal beam 110. Imaging optics 114, e.g., a Fourier-transforming lens, relays the data from the SLM 112 to the holographic medium 102. The reference arm 106 has a motor-driven rotating mirror 122 and a 4f imaging system having lenses 124 and 126. The reference beam 120, often a plane-wave, is directed to the holographic medium 102 at a specific angle. The coherent interference between the data-bearing signal beam 110 and the reference beam 120 in the holographic medium 102 records one page of data corresponding to that specific angle. Multiple data pages are recorded and superimposed by changing the angle of the reference beam 120 relative to the motor-driven rotating mirror 122.
To access a specific data page within the memory during readout, the reference beam 120 is directed to the holographic medium 102 at the angle at which that data page is recorded. A reconstruction beam 130 was generated by the diffraction of the reference beam 120 from the hologram stored in the medium 102. Imaging optics 132 (e.g., a Fourier-transforming lens) and a detector array 134 (e.g., a CCD camera) are used to retrieve the data in the reconstruction beam 130. The motor-driven mirror 122 directs the reference beam 120 to address individual data pages within the memory.
The above described system 100 is a holographic memory based on angularly multiplexed volume holograms. Such a structure is well known in the art. Each data page represents a two dimensional data array. Many data pages can be holographically multiplexed into the same recording volume. The holographic medium 102 can be made very compact in size. A storage density of 100 GB per cm.sup.3 is possible. Other holographic multiplexing techniques can be used to superimpose multiple data pages in a common recording volume including phase-code multiplexing, fractal multiplexing and peristrophic multiplexing, shift multiplexing, and wavelength multiplexing. The storage capacity of a volume multiplexed holographic memory of volume V is of the order of V/.lambda..sup.3 where .lambda. is the light wavelength. This number can be as large as 10.sup.12 for moderate crystal sizes (e.g. 1 cm), but usually factors such as dynamic range and optics apertures reduce it by several orders of magnitude to an estimate of several GB per cm.sup.3. As an example, 160,000 holograms were recently demonstrated in 1 cm of Fe-doped LiNbO.sub.3 by using angle and fractal multiplexing in the 90.degree. geometry, and 1,000 holograms were recorded in 100 .mu.m thick photopolymer by using angle and peristrophic multiplexing in the holographic disk geometry.
The data readout in such holographic memory systems is fast. A page is often read out in the form of a two dimensional array by the reference read beam 120 simultaneously in parallel at the speed of light. Each data page can be randomly accessed by simply directing the readout reference beam 120 at a proper angle.
However, lenses are required to image the reconstructed data page on the detector array 134 in the conventional system 100 in FIG. 1. The required separation between the lenses, the SLM and the detector array imposes a minimum size on the system. The detector array 134 must be aligned to within a fraction of the pixel size with respect to the SLM 112 in order to properly register the multiplexed data pages in the readout. The motorized mirror 122, however, usually has significant inertia. This provides a high power dissipation and low response speed. The motorized mirror 122 also may have backlash which prevents Bragg matching the stored holograms correctly. The mechanical motion of the system thus significantly hinders the full potential of high speed of the holographic memory. In addition, such a system is difficult to integrate with other electronic data processing units.
The second example of conventional optical holographic system is an optical neural network. Artificial neural networks are usually arranged in layers. Neurons at each layer do not interact with neurons at the same layer, but are densely interconnected to neurons at other layers. In a typical implementation, shown in FIG. 2a, the neural network has two layers. The input layer 202 receives information (e.g. sensory data of spatial or temporal nature) in the form of a real vector (f.sub.1, . . . f.sub.n). The middle layer 204 performs a nonlinear operation g.sup.(1) to weighted sums of input components, producing the internal representation (h.sub.1, . . . , h.sub.m), where ##EQU1## The role of the middle layer 204 is to apply a nonlinear transformation to its input that will make the internal representation of input classes linearly separable. The output layer 206 (very often a single neuron) combines outputs of the hidden layers and applies a nonlinearity g.sup.(2) to the result before producing the classification (or decision) vector y.sub.i, i.e. ##EQU2## Networks with more than two layers can be constructed, but the two-layer architecture is most common, because it is sufficient for any segmentation of the input space with relatively cost-efficient training.
Holographic optical systems offer an elegant solution for the implementation of the interconnections in neural networks. In contrast to electrical signals, optical beams are not affected by capacitive coupling, therefore large numbers of interconnects may coexist in a small volume. They all operate at the speed of light or at some fraction thereof. The mathematical model for the neural network usually determines the optical implementation.
One optical implementation of the system of FIG. 2a is shown in FIG. 2b. Volume holograms stored in an appropriate recording medium define the interconnects. The strengths can be either precomputed and stored directly, or trained in real-time. Popular materials for this kind of architecture include photorefractives; in the case of disk-based implementations, other media such as photothermoplastic, optical memory disks, and photopolymers have been investigated. Since each hologram can carry 10.sup.5 -10.sup.6 bits (each bit corresponds to a single plane wave hologram component in the Fourier plane geometry), the capacity of these systems, expressed as simple neuron-to-neuron interconnects, can be as high as 10.sup.11.
Various ways of organizing interconnects are possible. FIG. 2b shows an implementation by an optical volume holographic correlator. The Fourier transforms of the training templates are holographically multiplexed and stored in the holographic medium 220. A test template is imprinted on a beam 210 by a SLM 212. The test template is Fourier transformed by a lens 214 and simultaneously correlated with all stored templates by reading out the multiplexed holograms in the holographic medium 220. An output reconstruction beam 222 is generated. The Fourier transforms of the reconstructed references by a lens 224 yield the cross-correlations, which are captured by an array of photodetectors 226 in the output plane.
The degree of similarity with the stored patterns in the holographic medium 220 is thus determined, and an electronic nonlinearity can perform, e.g. a thresholding operation to determine whether the template presented to the correlator belongs to the set of stored templates. This method has been used successfully for pattern recognition such as face and fingerprint recognition. A more complicated output layer algorithm was implemented for target tracking and the indoor navigation of a robot car.
The correlator geometry described above has many advantages in addition to the capability of large number of interconnects. For example, one advantage is the invariance of the correlation to translation in the input pattern, resulting from correlating in the Fourier domain. When three-dimensional materials are used, Bragg selectivity limits the shift invariance in one dimension (in the plane defined by the reference and the signal carrier), but the remaining vertical shift invariance is sufficient to allow successful generalization in many systems. Another advantage is that the stored templates can be directly recalled in the same setup if read-out by the reference beams, typically one at a time.
Many other optical systems can be built based on the system in FIG. 2b. For example, one can construct an associative memory which is a simple model for a neural network capable of recalling one out of several stored patterns when a distorted version of the desired pattern is presented at its input.
However, the system in FIG. 2b, also suffers limitations. For example, since a holographic medium 220 is used to store data-bearing holograms, the conventional imaging optical systems and beam steering elements are used for recording and reconstructing the holograms. This presents serious obstacles to achieving a high performance and compact neural network. In addition, the bulky conventional systems are expensive to manufacture. Despite the distinct advantages and versatile applications, the conventional optical components and architecture are limiting in full utilization of these advantages.
In recognition of the above, the inventors of the present invention developed new system architectures based on volume holographic storage. The present invention will improve the compactness, ease of alignment, cost, weight and suitability to industrial fabrication of random-access volume holographic rewritable memories and other optoelectronic systems. The new system architectures use modules having three dimensional holographic materials, programmable beam steering devices such as diffractive optical elements, and optoelectronic integrated circuits (OEIC) using modulating elements such as liquid crystal modulators for light modulation.
The conventional complex imaging optical elements such as 4f lens assembly are eliminated from the signal arm in the new system architectures in accordance with the present invention. This is accomplished in the preferred embodiments by using integrated component interconnections and conjugate readout. This presents several advantages. For example, the difficulty in achieving and maintaining the desired alignment is significantly reduced. The adverse effects caused by factors such as vibrations are thus minimized. Elimination of imaging optics further reduces the size of the system and facilitates system integration. In particular, phase conjugators are used to implement the conjugate readout. Phase conjugation yields beams that retrace themselves, thereby canceling any aberrations present in the backward path. Thus the need for imaging optics is eliminated.
The preferred embodiments of the present invention use programmable diffractive optical elements to accurately control the directions of optical beams for recording and retrieving data and for information processing operations. According to the present invention, liquid crystal deflectors are integrated with the holographic cubes to replace the motor-driven mirror or acoustic-opto modulators along with imaging optical elements used in many conventional systems for redirecting light beams. This largely eliminates the problems associated with conventional mechanical addressing systems (e.g., vibrations, backlash, low speed, high power consumption) and with acousto-optic deflectors (e.g., wavelength shift, high-voltage operation). This reduces the dimension of the systems, increases the response speed, simplifies the optical alignment, and enhances the system robustness.
One unique aspect of the present invention is a multi-functional OEIC that integrates a modulator array and a detector array together to form a spatial light modulator and a two-dimensional detector array on the some surface of a die. The conjugate readout used in the preferred embodiments is self-retracing. The compact detectors and light modulators can be spatially close to each other or physically superimposed together. One significant advantage of such integration is that the alignment of the detectors and the modulators is automatically guaranteed. The stringent aligning requirement of the detector array relative to the SLM that is required in a conventional holographic system is hence significantly relaxed. Also, the cost of commercial fabrication of holographic systems is greatly reduced since well-developed semiconductor fabrication processes can be used in making the OEICs with high yield.
In operation, the OEIC of the present invention can be programmed to perform spatial light modulation fcr recording two dimensional data pages or to detect light signals for retrieving data. Furthermore, it can also periodically refresh the dynamic holograms that slowly decay as a result of accesses to the read/write memory. Thus, the OEIC of the present invention is also a Dynamic Holographic Refresher (DHR).
Volume multiplexed holograms are implemented in the integrated architectures in accordance with the present invention. Angularly multiplexed module, wavelength multiplexed module, phase-coded module, and shift-multiplexed module are the preferred embodiments of the present invention.
The new integrated system architectures have versatile applications in a number of optical holographic systems including, but not limited to, high-density holographic memory, holographic correlators for image recognition, holographic associative memories, and holographic interconnects.